Array API support

cuPyNumeric’s Array API support is incremental and is not a conformance claim. It exposes the namespace dispatch and inspection hooks used by Array API consumers, along with selected standard function spellings backed by native cuPyNumeric implementations.

cuPyNumeric currently exposes None as its only Array API device token. None here means Legate-managed placement, not CPU placement. Concrete device strings such as "cpu" or "cuda:0" are rejected by the Array API hooks.

Array API functions with wrapper behavior are listed here. Direct Array API aliases are listed in the matching functional API sections.

cupynumeric.__array_api_version__

The Array API standard version implemented by cuPyNumeric’s Array API namespace.

__array_namespace_info__()

Return Array API namespace inspection utilities.

astype(x, dtype, /, *[, copy, device])	Cast an array to an Array API dtype.
isdtype(dtype, kind, /)	Return whether a dtype matches an Array API dtype kind.
matrix_transpose(x, /)	Transpose the last two dimensions of an array.
unique_all(x, /)	Find unique values and their indices, inverse indices, and counts.
unique_counts(x, /)	Find unique values and their occurrence counts.
unique_inverse(x, /)	Find unique values and the indices needed to reconstruct the input.
unique_values(x, /)	Find the unique values in an array.
vecdot(x1, x2, /, *[, axis])	Compute the vector dot product along an axis.

class cupynumeric.ArrayNamespaceInfo

Array API namespace inspection utilities.

Methods

capabilities()
default_device()
default_dtypes(*[, device])
devices()
dtypes(*[, device, kind])